A long time tracking with BIN-NST and DRN

Abstract

Long time tracking in video stream is one of the most challenging problem in computer vision. Most state-of-the-art visual trackers rely on the classical sliding windows, resulting in a large number of candidate object window. In order to achieve purpose of real-time tracking, the weak feature is used for object description. So in many complex scenes, such as illumination variation, the algorithm performance is seriously decreased. Firstly, a more efficient object detection algorithm called BIN-NST is used to improve the detector in the tracking–learning–detection framework. We find that the normed gradients, designed for generic objectness estimation, are also able to rapidly generate high quality candidate windows. We also notice that using the tensor feature to replace the gradients feature can improve the performance of objectness estimation. Based on these observations, we propose an efficient method, which we call BIN-NST, to produce candidate windows. Experiments on the challenging PASCAL VOC 2007 dataset show that our method efficiently generates a small set of category-independent, high quality object windows. The performance of BIN-NST and BING is almost the same in the single scene, but the performance of BIN-NST in the complex scene is significantly higher than that of the BING. In addition, compared to the classical sliding windows mechanism used in the tracking–learning–detection framework, the candidate window generated by BIN-NST is reduced by about 90%. Secondly, GOTURN tracker is a method for offline training of neural networks that can track novel objects at test-time at 100 fps. The GOTURN tracker is significantly faster than previous methods that use neural networks for tracking. But the GOTURN tracker will fail to track object when the object become occluded and is moving too quickly. In this paper, these candidate object windows, which are detected by object detector based on BIN-NST, are fed to the network of GOTURN. The object extraction ability of the detector is used to solve the problem that the GOTURN algorithm cannot track fast moving object effectively. The experiment shows that the tracking algorithm can still track the object effectively when the object is moving quickly.